Semi-supervised Adversarial Learning for Diabetic Retinopathy Screening

Abstract

It is well known that in medical image analysis, only a small number of high-quality labeled images can be often obtained from a large number of medical images due to the requirement of expert knowledge and intensive labor work. Therefore, we propose a novel semi-supervised adversarial learning framework (SSALF) for diabetic retinopathy (DR) screening of color fundus images. Specifically, our proposed framework consists of two subnetworks, an extended network and a discriminator. The extended network is obtained by extending a common classification network with a generator used for unsupervised image reconstruction. Thus, the extended network can utilize some labeled and lots of unlabeled fundus images. Then the discriminator is attached to the generator of the extended network to judge whether a reconstructed image is real or fake, introducing adversarial learning into the whole framework. Our framework achieves promising utility and generalization on the datasets of EyePACS and Messidor in a semi-supervised setting: we use some labeled and lots of unlabeled fundus images to train our framework. And we also investigate the effects of image reconstruction and adversarial learning on our framework by implementing ablation experiments.